Waymo built its own chip for robotaxis, cutting reliance on Nvidia

Waymo, Alphabet's robotaxi subsidiary, built its own custom chip for its self-driving taxis, cutting its reliance on Nvidia, The Decoder reported, citing Bloomberg. The chip is already running in the company's latest generation of vehicles.

Waymo says the chip processes sensor data faster and runs the AI models that help its robotaxis understand and react to their surroundings. Until now, the company relied on chips from Nvidia and AMD. The custom chip delivers more than 1,000 TOPS, short for trillions of operations per second, and is manufactured by TSMC on a 5-nanometer process, according to Waymo. That puts it on par with Nvidia's current autonomous driving systems.

TOPS matters as a metric because a robotaxi has to fuse camera, lidar, and radar data in real time, then run perception, prediction, and planning models fast enough to react to traffic before the moment passes. All of that has to happen inside a vehicle's strict power and latency budget, unlike a data center where a chip can draw far more power without the same physical constraints. A higher TOPS number at a given power draw means the chip can process more of that sensor data per second without pushing the car's energy and cooling systems past their limits.

• Chip performance: more than 1,000 TOPS
• Manufacturing: TSMC, 5-nanometer process
• Previous suppliers: Nvidia and AMD
• Deployment: already running in Waymo's newest vehicles
• New robotaxi partner: Chinese manufacturer Zeekr

Building custom silicon, a chip designed for one company's specific workload rather than sold as general-purpose hardware, tends to make sense only once that company's software stack and production volume are mature enough to justify the design cost. A general-purpose GPU has to handle a wide range of workloads well, which leaves performance and efficiency on the table for any single application. A chip built specifically for Waymo's perception and planning software can strip out anything it doesn't need and tune the remaining silicon for that exact job, which is where the cost savings Waymo is pointing to come from.

The move fits a broader pattern across AI-heavy companies: renting or buying merchant silicon while a product is young, then designing proprietary chips once scale justifies the investment. A similar dynamic runs the opposite direction in how Meta pays Microsoft hundreds of millions for Azure AI access even while building its own competing infrastructure, the same build-your-own-versus-depend-on-a-supplier tension showing up on both sides of the AI hardware market.

The chip shift also carries strategic weight for Waymo's cost structure. Robotaxi fleets run compute in every vehicle around the clock, so shaving power draw and hardware cost per car compounds fast at scale in a way it wouldn't for a single test vehicle. Nvidia and AMD chips are priced for a broad market of customers with different needs, and Waymo paying for capabilities it doesn't use adds up across thousands of vehicles. A chip stripped down to exactly what Waymo's software needs removes that overhead, which is the direct source of the cost reduction the company is pointing to.

Waymo is installing the new chip in its next robotaxi, which it's building with Chinese manufacturer Zeekr, marking the first vehicle where the company's custom hardware and a new international manufacturing partner ship together.

This piece is informational, not a recommendation to buy, sell, or hold any asset.